Pretrained Weights

keras_climate.weights.pretrained provides "timm-style" pretrained-weight loaders: each function downloads a real, publicly hosted PyTorch checkpoint (cached locally so it's only fetched once), ports it through WeightConverter + the matching name-mapping, and returns a ready keras.Model (or encoder/decoder pair) with real pretrained weights loaded — no training required to get a working model.

from keras_climate.weights.pretrained import unet_carvana

model, report = unet_carvana()
model.summary()

Every loader returns (model, report) (or (encoder, decoder, report) for SatMAE) where report is the WeightConverter.convert(...) report dict, so callers can inspect exactly what did/didn't get matched.

Coverage

Not every architecture in this repo has a publicly downloadable checkpoint that matches it exactly.

Loader	Checkpoint	Match
unet_carvana	milesial/Pytorch-UNet's Carvana release	exact architecture + checkpoint
deeplabv3plus_resnet50_imagenet_backbone	torchvision ImageNet-1k ResNet-50	backbone only — no public ASPP/decoder checkpoint exists
segformer_b0_ade20k	nvidia/segformer-b0-finetuned-ade-512-512 (HF)	exact — requires transformers
satmae_vit_base_mae	Meta AI's official ImageNet-1k MAE ViT-Base	stand-in — SatMAE's own checkpoints aren't hosted; mode="single" is architecturally identical to plain ViT-MAE
prithvi_eo_100m	ibm-nasa-geospatial/Prithvi-EO-1.0-100M	exact — encoder only
croma_base / croma_large	antofuller/CROMA (CROMA_base.pt / CROMA_large.pt)	exact — both modalities + joint fusion
scalemae_vitlarge_fmow	TorchGeo re-export of facebookresearch/scale-mae fMoW-RGB ViT-L/16	exact — decoder/FPN not implemented
ssl4eo_resnet50_moco	TorchGeo re-export of zhu-xlab/SSL4EO-S12 MoCo v2 ResNet-50	exact
satclip_location_encoder_resnet18_l10	Microsoft's official SatCLIP-ResNet18-L10	exact — location encoder only, not the paired image encoder
prithvi_eo_v2_300m	ibm-nasa-geospatial/Prithvi-EO-2.0-300M	exact — encoder only, base non-"-TL" variant
fourcastnet_backbone	NVIDIA's official release, mirrored at NERSC	exact — requires ruamel.yaml + torch
climax_1_40625deg	Microsoft's official microsoft/ClimaX (1.40625deg.ckpt)	exact
pangu_weather_24	Clean PyTorch conversion of Huawei's ONNX release	exact — BY-NC-SA 4.0, non-commercial use only
anysat_base	g-astruc/AnySat (HuggingFace), "base" config	exact — validated end-to-end to ~1e-6 against real PyTorch forward passes

No loader available

• AFNO — the only released AFNO weights are FourCastNet's own (already covered by fourcastnet_backbone); no standalone, task-agnostic AFNO checkpoint exists.
• Clay — the official checkpoint is ~5GB, impractical to fetch/validate in most environments; still validated against a synthetic reference matching the assumed mapping naming, just not the real public checkpoint.
• AnySatEncoder (this repo's own simpler AnySat design) — validated only against a from-scratch synthetic reference, not the real checkpoint (see anysat_base/AnySatRelease above for the faithful port that is checkpoint-compatible).
• RingMo — no official or credible unofficial checkpoint has ever been publicly released for it at all.

All unloaded models are validated only against from-scratch synthetic PyTorch references of their own architectures — see Weight Porting.

Out of scope

FourCastNet v2 (SFNO) is deliberately not implemented: it replaces AFNO's planar FFT with a genuine Spherical Harmonic Transform on the sphere, which is the model's defining feature (not an implementation detail) and would require reproducing a correctness-critical numerical transform library (torch-harmonics) from scratch to load its real checkpoint faithfully.